Research
Twelve research
directions
We work in twelve directions, from the numerical core to applied systems. Each one below opens with a short summary and one representative paper.
Our current focus is small models, interpretability of LLMs, new architectures, scientific machine learning, and applications. Much of this work builds on one family of methods: low-rank approximations, which replace a large matrix or tensor with a product of much smaller ones. The same approximations compress a neural network, accelerate a PDE solver, and rank recommendations.
The numerical core of the lab. We develop the tensor-train format and related low-rank tools for high-dimensional problems: black-box approximation, quantized tensor FEM, and open-source software used across numerical analysis and quantum chemistry.
Tensor Train Decomposition on TensorFlow (T3F) · JMLR 2020We make deep models smaller and faster with low-rank and tensor factorizations: compressed convolutional and embedding layers, tensorized pre-training, faster LoRA fine-tuning, and expert placement for Mixture-of-Experts inference.
Run LoRA Run: Faster and Lighter LoRA Implementations · ACL 2025We probe the internal structure of transformers: near-linear layer maps, sparse autoencoders for reasoning features, the role of punctuation in context memory, task vectors for in-context learning, and hidden capacity for one-step generation.
Your Transformer is Secretly Linear · ACL 2024We study attacks and defenses together: universal adversarial perturbations, certified robustness of speech and vision models, membership inference, machine unlearning in multimodal models, and moderation of LLM outputs.
CLEAR: Character Unlearning in Textual and Visual Modalities · Findings of ACL 2025We fuse neural operators with classical iterative methods: learned preconditioners via graph networks, physics-informed and spectral-informed models, Kolmogorov-Arnold networks for singular PDEs, and data-driven weather forecasting.
Neural operators meet conjugate gradients: FCG-NO for PDE solving · ICML 2024We build gradient-free optimizers on tensor-train sampling and maximum-volume submatrices, and apply them to reinforcement learning, activation maximization, and global search over atomic clusters.
PROTES: Probabilistic Optimization with Tensor Sampling · NeurIPS 2023We run evolutionary loops where an LLM acts as the mutation and crossover operator over programs, in the spirit of AlphaEvolve. Our open-source framework GigaEvo carries this from quantum circuit optimization to medical pipelines.
GigaEvo: An Open Source Optimization Framework Powered by LLMs and Evolution · arXiv 2025We assemble systems from cooperating LLM agents: text-to-Cypher GraphRAG over labeled property graphs, and clinical platforms where agents route patients, infer conditions, and triage inside structured dialogues.
CLARITY: Clinical Assistant for Routing, Inference, and Triage · EMNLP 2025 IndustryWe work on generative models from theory to practice: GAN convergence, optimal-transport analysis of diffusion latent codes, fusion of diffusion adapters, discrete diffusion language models, and Riemannian flow matching for molecular docking.
Understanding DDPM Latent Codes Through Optimal Transport · ICLR 2023We have built recommenders for a decade: tensor- and SVD-based collaborative filtering, negative feedback modeling, hyperbolic representations, and scalable losses for sequential recommenders with large item catalogs.
Scalable Cross-Entropy Loss for Sequential Recommendations · RecSys 2024We take representation learning beyond Euclidean space: hyperbolic embeddings for images, metric learning with hyperbolic vision transformers, hyperbolic recommenders, and Riemannian optimization on low-rank manifolds.
Hyperbolic Image Embeddings · CVPR 2020We bring these methods to practice: clinical decision support and triage built on LLMs, radiography report generation, robotics, remote sensing, quantum circuit simulation, and speech recognition of mathematical formulas.
AI Diagnostic Assistant (AIDA) for Clinical Decision Support · IJCAI 2025